diff --git a/.github/workflows/claude-issue-investigate.yml b/.github/workflows/claude-issue-investigate.yml
new file mode 100644
index 000000000..0189701cf
--- /dev/null
+++ b/.github/workflows/claude-issue-investigate.yml
@@ -0,0 +1,75 @@
+name: Investigate issue
+
+# Gated on a maintainer-added label rather than issues: opened, since this is a public repo and anyone can open an issue - this way a human triages first before spending budget on a Claude session.
+on:
+  issues:
+    types: [labeled]
+
+permissions:
+  contents: write
+  pull-requests: write
+  issues: read
+  id-token: write
+
+jobs:
+  investigate:
+    if: github.event.label.name == 'investigate'
+    runs-on: ubuntu-latest-large
+    env:
+      ARTIFACTORY_URL: ${{ vars.ARTIFACTORY_URL }}
+    steps:
+      - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
+
+      - name: Artifactory OIDC Auth
+        uses: ./.github/actions/artifactory-oidc
+
+      - uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4
+        with:
+          node-version: 22
+
+      - run: corepack enable
+
+      - name: Install dependencies
+        run: yarn install --immutable
+
+      - name: Install Claude Code CLI
+        run: npm install -g @anthropic-ai/claude-code
+
+      # New commits need an author - Claude never has repo write access outside this job.
+      - name: Configure git identity for Claude's commits
+        run: |
+          git config --global user.email "claude-bot@users.noreply.github.com"
+          git config --global user.name "claude-bot"
+
+      # ISSUE_NUMBER only - title/body are untrusted, so Claude fetches them itself via gh rather than us splicing them into this script.
+      - name: Write prompt
+        env:
+          ISSUE_NUMBER: ${{ github.event.issue.number }}
+        run: |
+          cat > "$RUNNER_TEMP/claude-prompt.txt" <<EOF
+          Investigate GitHub issue #$ISSUE_NUMBER in this repository.
+
+          1. Run \`gh issue view $ISSUE_NUMBER\` to read the full issue (title, body, comments).
+          2. Based on the description, actually try to reproduce the reported problem using this repository's existing code and test tooling - write and run a small script or test case that exercises the described behavior, don't just reason about whether it would fail.
+          3. If you reproduce it, identify the root cause in the source code.
+          4. If you can reproduce it and you're confident the fix is small, narrow, and clearly correct, implement it, create a new branch off master named \`fix/issue-$ISSUE_NUMBER\`, commit, push it, and open a pull request against master with \`gh pr create\` that includes "Fixes #$ISSUE_NUMBER" in the body. If the right fix is unclear, would need broader changes, or you're not confident it's correct, do not open a PR - note that in the report instead.
+          5. If you cannot reproduce it, explain exactly what you tried, why it didn't reproduce, and what additional information (exact repro steps, versions, environment) would help.
+          6. Do not comment on the issue itself.
+          7. Always write your findings to investigation-report.md in the repository root: a summary, the exact reproduction steps you tried, whether it reproduced, your root-cause diagnosis if found, and whether you opened a fix PR (with its number) or why you decided not to.
+          EOF
+
+      - name: Investigate with Claude
+        env:
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          claude -p "$(cat "$RUNNER_TEMP/claude-prompt.txt")" \
+            --allowedTools "Edit,MultiEdit,Write,Read,Glob,Grep,Bash(git:*),Bash(gh:*),Bash(yarn:*),Bash(npm:*),Bash(node:*)"
+
+      - name: Upload investigation report
+        if: always()
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4
+        with:
+          name: issue-${{ github.event.issue.number }}-investigation
+          path: investigation-report.md
+          if-no-files-found: warn